Research Of Object Detection Methods For Incomplete Sample Annotations

As a critical problem in the field of computer vision,object detection has been extensively utilized in diverse domains such as face recognition,power system detection,and remote sensing image processing.With the improvement of computer computing power and the in-depth research of deep learning algorithms,supervised learning methods have achieved excellent learning results,but this relies on large-scale annotated samples behind it.Object detection needs to accomplish both recognition and localization.Relying solely on experts to annotate all samples is a very time-consuming and laborious process,posing significant challenges in terms of manpower and resources.Moreover,in practical research,some application scenarios cannot meet the requirements of complete annotation.Therefore,how to conduct semi-supervised learning based on limited annotated samples and how to annotate more valuable samples under the limited annotated resources have become significant research questions with practical significance in this paper.This paper takes the insulator in high-voltage power systems as the primary research object and introduces related research theories such as deep learning and computer vision to analyze and model the problem of semi-supervised object detection under limited annotation.The main work and research achievements are summarized as follows:1.An object detection framework was proposed to address the issue of incomplete sample annotation by incorporating PU(Positive-Unlabeled)learning into the Faster RCNN network architecture.We optimized the original estimation method of PU learning and evaluated the proposed method on the Insulator Defect Image Dataset(IDID).The results demonstrate the effectiveness of our approach in scenarios with incomplete label annotation,and we compared the experimental performance of our method with several different mainstream object detection backbone networks.2.In the context of limited expert annotation resources for object detection in industrial inspection,a PU-AL(Positive Unlabeled-Active Learning)joint training framework was proposed in our work,which combines Active learning with PU learning to address the problem of incomplete annotations in the dataset.Our approach integrates the Active learning method into the PU learning framework by introducing a loss branch prediction network to evaluate the information content of unlabeled samples,and selectively annotates samples based on the predicted loss value.This enables the network to achieve better training performance with the same annotation cost.The training method and detailed annotation strategy of the proposed joint training network was given in this paper.We also compare the experimental results of randomly annotating samples with those of our annotation strategy.